...

Source file src/crypto/internal/bigmod/nat.go

Documentation: crypto/internal/bigmod

     1  // Copyright 2021 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package bigmod
     6  
     7  import (
     8  	"errors"
     9  	"internal/byteorder"
    10  	"math/big"
    11  	"math/bits"
    12  )
    13  
    14  const (
    15  	// _W is the size in bits of our limbs.
    16  	_W = bits.UintSize
    17  	// _S is the size in bytes of our limbs.
    18  	_S = _W / 8
    19  )
    20  
    21  // choice represents a constant-time boolean. The value of choice is always
    22  // either 1 or 0. We use an int instead of bool in order to make decisions in
    23  // constant time by turning it into a mask.
    24  type choice uint
    25  
    26  func not(c choice) choice { return 1 ^ c }
    27  
    28  const yes = choice(1)
    29  const no = choice(0)
    30  
    31  // ctMask is all 1s if on is yes, and all 0s otherwise.
    32  func ctMask(on choice) uint { return -uint(on) }
    33  
    34  // ctEq returns 1 if x == y, and 0 otherwise. The execution time of this
    35  // function does not depend on its inputs.
    36  func ctEq(x, y uint) choice {
    37  	// If x != y, then either x - y or y - x will generate a carry.
    38  	_, c1 := bits.Sub(x, y, 0)
    39  	_, c2 := bits.Sub(y, x, 0)
    40  	return not(choice(c1 | c2))
    41  }
    42  
    43  // Nat represents an arbitrary natural number
    44  //
    45  // Each Nat has an announced length, which is the number of limbs it has stored.
    46  // Operations on this number are allowed to leak this length, but will not leak
    47  // any information about the values contained in those limbs.
    48  type Nat struct {
    49  	// limbs is little-endian in base 2^W with W = bits.UintSize.
    50  	limbs []uint
    51  }
    52  
    53  // preallocTarget is the size in bits of the numbers used to implement the most
    54  // common and most performant RSA key size. It's also enough to cover some of
    55  // the operations of key sizes up to 4096.
    56  const preallocTarget = 2048
    57  const preallocLimbs = (preallocTarget + _W - 1) / _W
    58  
    59  // NewNat returns a new nat with a size of zero, just like new(Nat), but with
    60  // the preallocated capacity to hold a number of up to preallocTarget bits.
    61  // NewNat inlines, so the allocation can live on the stack.
    62  func NewNat() *Nat {
    63  	limbs := make([]uint, 0, preallocLimbs)
    64  	return &Nat{limbs}
    65  }
    66  
    67  // expand expands x to n limbs, leaving its value unchanged.
    68  func (x *Nat) expand(n int) *Nat {
    69  	if len(x.limbs) > n {
    70  		panic("bigmod: internal error: shrinking nat")
    71  	}
    72  	if cap(x.limbs) < n {
    73  		newLimbs := make([]uint, n)
    74  		copy(newLimbs, x.limbs)
    75  		x.limbs = newLimbs
    76  		return x
    77  	}
    78  	extraLimbs := x.limbs[len(x.limbs):n]
    79  	clear(extraLimbs)
    80  	x.limbs = x.limbs[:n]
    81  	return x
    82  }
    83  
    84  // reset returns a zero nat of n limbs, reusing x's storage if n <= cap(x.limbs).
    85  func (x *Nat) reset(n int) *Nat {
    86  	if cap(x.limbs) < n {
    87  		x.limbs = make([]uint, n)
    88  		return x
    89  	}
    90  	clear(x.limbs)
    91  	x.limbs = x.limbs[:n]
    92  	return x
    93  }
    94  
    95  // set assigns x = y, optionally resizing x to the appropriate size.
    96  func (x *Nat) set(y *Nat) *Nat {
    97  	x.reset(len(y.limbs))
    98  	copy(x.limbs, y.limbs)
    99  	return x
   100  }
   101  
   102  // setBig assigns x = n, optionally resizing n to the appropriate size.
   103  //
   104  // The announced length of x is set based on the actual bit size of the input,
   105  // ignoring leading zeroes.
   106  func (x *Nat) setBig(n *big.Int) *Nat {
   107  	limbs := n.Bits()
   108  	x.reset(len(limbs))
   109  	for i := range limbs {
   110  		x.limbs[i] = uint(limbs[i])
   111  	}
   112  	return x
   113  }
   114  
   115  // Bytes returns x as a zero-extended big-endian byte slice. The size of the
   116  // slice will match the size of m.
   117  //
   118  // x must have the same size as m and it must be reduced modulo m.
   119  func (x *Nat) Bytes(m *Modulus) []byte {
   120  	i := m.Size()
   121  	bytes := make([]byte, i)
   122  	for _, limb := range x.limbs {
   123  		for j := 0; j < _S; j++ {
   124  			i--
   125  			if i < 0 {
   126  				if limb == 0 {
   127  					break
   128  				}
   129  				panic("bigmod: modulus is smaller than nat")
   130  			}
   131  			bytes[i] = byte(limb)
   132  			limb >>= 8
   133  		}
   134  	}
   135  	return bytes
   136  }
   137  
   138  // SetBytes assigns x = b, where b is a slice of big-endian bytes.
   139  // SetBytes returns an error if b >= m.
   140  //
   141  // The output will be resized to the size of m and overwritten.
   142  func (x *Nat) SetBytes(b []byte, m *Modulus) (*Nat, error) {
   143  	if err := x.setBytes(b, m); err != nil {
   144  		return nil, err
   145  	}
   146  	if x.cmpGeq(m.nat) == yes {
   147  		return nil, errors.New("input overflows the modulus")
   148  	}
   149  	return x, nil
   150  }
   151  
   152  // SetOverflowingBytes assigns x = b, where b is a slice of big-endian bytes.
   153  // SetOverflowingBytes returns an error if b has a longer bit length than m, but
   154  // reduces overflowing values up to 2^⌈log2(m)⌉ - 1.
   155  //
   156  // The output will be resized to the size of m and overwritten.
   157  func (x *Nat) SetOverflowingBytes(b []byte, m *Modulus) (*Nat, error) {
   158  	if err := x.setBytes(b, m); err != nil {
   159  		return nil, err
   160  	}
   161  	leading := _W - bitLen(x.limbs[len(x.limbs)-1])
   162  	if leading < m.leading {
   163  		return nil, errors.New("input overflows the modulus size")
   164  	}
   165  	x.maybeSubtractModulus(no, m)
   166  	return x, nil
   167  }
   168  
   169  // bigEndianUint returns the contents of buf interpreted as a
   170  // big-endian encoded uint value.
   171  func bigEndianUint(buf []byte) uint {
   172  	if _W == 64 {
   173  		return uint(byteorder.BeUint64(buf))
   174  	}
   175  	return uint(byteorder.BeUint32(buf))
   176  }
   177  
   178  func (x *Nat) setBytes(b []byte, m *Modulus) error {
   179  	x.resetFor(m)
   180  	i, k := len(b), 0
   181  	for k < len(x.limbs) && i >= _S {
   182  		x.limbs[k] = bigEndianUint(b[i-_S : i])
   183  		i -= _S
   184  		k++
   185  	}
   186  	for s := 0; s < _W && k < len(x.limbs) && i > 0; s += 8 {
   187  		x.limbs[k] |= uint(b[i-1]) << s
   188  		i--
   189  	}
   190  	if i > 0 {
   191  		return errors.New("input overflows the modulus size")
   192  	}
   193  	return nil
   194  }
   195  
   196  // Equal returns 1 if x == y, and 0 otherwise.
   197  //
   198  // Both operands must have the same announced length.
   199  func (x *Nat) Equal(y *Nat) choice {
   200  	// Eliminate bounds checks in the loop.
   201  	size := len(x.limbs)
   202  	xLimbs := x.limbs[:size]
   203  	yLimbs := y.limbs[:size]
   204  
   205  	equal := yes
   206  	for i := 0; i < size; i++ {
   207  		equal &= ctEq(xLimbs[i], yLimbs[i])
   208  	}
   209  	return equal
   210  }
   211  
   212  // IsZero returns 1 if x == 0, and 0 otherwise.
   213  func (x *Nat) IsZero() choice {
   214  	// Eliminate bounds checks in the loop.
   215  	size := len(x.limbs)
   216  	xLimbs := x.limbs[:size]
   217  
   218  	zero := yes
   219  	for i := 0; i < size; i++ {
   220  		zero &= ctEq(xLimbs[i], 0)
   221  	}
   222  	return zero
   223  }
   224  
   225  // cmpGeq returns 1 if x >= y, and 0 otherwise.
   226  //
   227  // Both operands must have the same announced length.
   228  func (x *Nat) cmpGeq(y *Nat) choice {
   229  	// Eliminate bounds checks in the loop.
   230  	size := len(x.limbs)
   231  	xLimbs := x.limbs[:size]
   232  	yLimbs := y.limbs[:size]
   233  
   234  	var c uint
   235  	for i := 0; i < size; i++ {
   236  		_, c = bits.Sub(xLimbs[i], yLimbs[i], c)
   237  	}
   238  	// If there was a carry, then subtracting y underflowed, so
   239  	// x is not greater than or equal to y.
   240  	return not(choice(c))
   241  }
   242  
   243  // assign sets x <- y if on == 1, and does nothing otherwise.
   244  //
   245  // Both operands must have the same announced length.
   246  func (x *Nat) assign(on choice, y *Nat) *Nat {
   247  	// Eliminate bounds checks in the loop.
   248  	size := len(x.limbs)
   249  	xLimbs := x.limbs[:size]
   250  	yLimbs := y.limbs[:size]
   251  
   252  	mask := ctMask(on)
   253  	for i := 0; i < size; i++ {
   254  		xLimbs[i] ^= mask & (xLimbs[i] ^ yLimbs[i])
   255  	}
   256  	return x
   257  }
   258  
   259  // add computes x += y and returns the carry.
   260  //
   261  // Both operands must have the same announced length.
   262  func (x *Nat) add(y *Nat) (c uint) {
   263  	// Eliminate bounds checks in the loop.
   264  	size := len(x.limbs)
   265  	xLimbs := x.limbs[:size]
   266  	yLimbs := y.limbs[:size]
   267  
   268  	for i := 0; i < size; i++ {
   269  		xLimbs[i], c = bits.Add(xLimbs[i], yLimbs[i], c)
   270  	}
   271  	return
   272  }
   273  
   274  // sub computes x -= y. It returns the borrow of the subtraction.
   275  //
   276  // Both operands must have the same announced length.
   277  func (x *Nat) sub(y *Nat) (c uint) {
   278  	// Eliminate bounds checks in the loop.
   279  	size := len(x.limbs)
   280  	xLimbs := x.limbs[:size]
   281  	yLimbs := y.limbs[:size]
   282  
   283  	for i := 0; i < size; i++ {
   284  		xLimbs[i], c = bits.Sub(xLimbs[i], yLimbs[i], c)
   285  	}
   286  	return
   287  }
   288  
   289  // Modulus is used for modular arithmetic, precomputing relevant constants.
   290  //
   291  // Moduli are assumed to be odd numbers. Moduli can also leak the exact
   292  // number of bits needed to store their value, and are stored without padding.
   293  //
   294  // Their actual value is still kept secret.
   295  type Modulus struct {
   296  	// The underlying natural number for this modulus.
   297  	//
   298  	// This will be stored without any padding, and shouldn't alias with any
   299  	// other natural number being used.
   300  	nat     *Nat
   301  	leading int  // number of leading zeros in the modulus
   302  	m0inv   uint // -nat.limbs[0]⁻¹ mod _W
   303  	rr      *Nat // R*R for montgomeryRepresentation
   304  }
   305  
   306  // rr returns R*R with R = 2^(_W * n) and n = len(m.nat.limbs).
   307  func rr(m *Modulus) *Nat {
   308  	rr := NewNat().ExpandFor(m)
   309  	n := uint(len(rr.limbs))
   310  	mLen := uint(m.BitLen())
   311  	logR := _W * n
   312  
   313  	// We start by computing R = 2^(_W * n) mod m. We can get pretty close, to
   314  	// 2^⌊log₂m⌋, by setting the highest bit we can without having to reduce.
   315  	rr.limbs[n-1] = 1 << ((mLen - 1) % _W)
   316  	// Then we double until we reach 2^(_W * n).
   317  	for i := mLen - 1; i < logR; i++ {
   318  		rr.Add(rr, m)
   319  	}
   320  
   321  	// Next we need to get from R to 2^(_W * n) R mod m (aka from one to R in
   322  	// the Montgomery domain, meaning we can use Montgomery multiplication now).
   323  	// We could do that by doubling _W * n times, or with a square-and-double
   324  	// chain log2(_W * n) long. Turns out the fastest thing is to start out with
   325  	// doublings, and switch to square-and-double once the exponent is large
   326  	// enough to justify the cost of the multiplications.
   327  
   328  	// The threshold is selected experimentally as a linear function of n.
   329  	threshold := n / 4
   330  
   331  	// We calculate how many of the most-significant bits of the exponent we can
   332  	// compute before crossing the threshold, and we do it with doublings.
   333  	i := bits.UintSize
   334  	for logR>>i <= threshold {
   335  		i--
   336  	}
   337  	for k := uint(0); k < logR>>i; k++ {
   338  		rr.Add(rr, m)
   339  	}
   340  
   341  	// Then we process the remaining bits of the exponent with a
   342  	// square-and-double chain.
   343  	for i > 0 {
   344  		rr.montgomeryMul(rr, rr, m)
   345  		i--
   346  		if logR>>i&1 != 0 {
   347  			rr.Add(rr, m)
   348  		}
   349  	}
   350  
   351  	return rr
   352  }
   353  
   354  // minusInverseModW computes -x⁻¹ mod _W with x odd.
   355  //
   356  // This operation is used to precompute a constant involved in Montgomery
   357  // multiplication.
   358  func minusInverseModW(x uint) uint {
   359  	// Every iteration of this loop doubles the least-significant bits of
   360  	// correct inverse in y. The first three bits are already correct (1⁻¹ = 1,
   361  	// 3⁻¹ = 3, 5⁻¹ = 5, and 7⁻¹ = 7 mod 8), so doubling five times is enough
   362  	// for 64 bits (and wastes only one iteration for 32 bits).
   363  	//
   364  	// See https://crypto.stackexchange.com/a/47496.
   365  	y := x
   366  	for i := 0; i < 5; i++ {
   367  		y = y * (2 - x*y)
   368  	}
   369  	return -y
   370  }
   371  
   372  // NewModulusFromBig creates a new Modulus from a [big.Int].
   373  //
   374  // The Int must be odd. The number of significant bits (and nothing else) is
   375  // leaked through timing side-channels.
   376  func NewModulusFromBig(n *big.Int) (*Modulus, error) {
   377  	if b := n.Bits(); len(b) == 0 {
   378  		return nil, errors.New("modulus must be >= 0")
   379  	} else if b[0]&1 != 1 {
   380  		return nil, errors.New("modulus must be odd")
   381  	}
   382  	m := &Modulus{}
   383  	m.nat = NewNat().setBig(n)
   384  	m.leading = _W - bitLen(m.nat.limbs[len(m.nat.limbs)-1])
   385  	m.m0inv = minusInverseModW(m.nat.limbs[0])
   386  	m.rr = rr(m)
   387  	return m, nil
   388  }
   389  
   390  // bitLen is a version of bits.Len that only leaks the bit length of n, but not
   391  // its value. bits.Len and bits.LeadingZeros use a lookup table for the
   392  // low-order bits on some architectures.
   393  func bitLen(n uint) int {
   394  	var len int
   395  	// We assume, here and elsewhere, that comparison to zero is constant time
   396  	// with respect to different non-zero values.
   397  	for n != 0 {
   398  		len++
   399  		n >>= 1
   400  	}
   401  	return len
   402  }
   403  
   404  // Size returns the size of m in bytes.
   405  func (m *Modulus) Size() int {
   406  	return (m.BitLen() + 7) / 8
   407  }
   408  
   409  // BitLen returns the size of m in bits.
   410  func (m *Modulus) BitLen() int {
   411  	return len(m.nat.limbs)*_W - int(m.leading)
   412  }
   413  
   414  // Nat returns m as a Nat. The return value must not be written to.
   415  func (m *Modulus) Nat() *Nat {
   416  	return m.nat
   417  }
   418  
   419  // shiftIn calculates x = x << _W + y mod m.
   420  //
   421  // This assumes that x is already reduced mod m.
   422  func (x *Nat) shiftIn(y uint, m *Modulus) *Nat {
   423  	d := NewNat().resetFor(m)
   424  
   425  	// Eliminate bounds checks in the loop.
   426  	size := len(m.nat.limbs)
   427  	xLimbs := x.limbs[:size]
   428  	dLimbs := d.limbs[:size]
   429  	mLimbs := m.nat.limbs[:size]
   430  
   431  	// Each iteration of this loop computes x = 2x + b mod m, where b is a bit
   432  	// from y. Effectively, it left-shifts x and adds y one bit at a time,
   433  	// reducing it every time.
   434  	//
   435  	// To do the reduction, each iteration computes both 2x + b and 2x + b - m.
   436  	// The next iteration (and finally the return line) will use either result
   437  	// based on whether 2x + b overflows m.
   438  	needSubtraction := no
   439  	for i := _W - 1; i >= 0; i-- {
   440  		carry := (y >> i) & 1
   441  		var borrow uint
   442  		mask := ctMask(needSubtraction)
   443  		for i := 0; i < size; i++ {
   444  			l := xLimbs[i] ^ (mask & (xLimbs[i] ^ dLimbs[i]))
   445  			xLimbs[i], carry = bits.Add(l, l, carry)
   446  			dLimbs[i], borrow = bits.Sub(xLimbs[i], mLimbs[i], borrow)
   447  		}
   448  		// Like in maybeSubtractModulus, we need the subtraction if either it
   449  		// didn't underflow (meaning 2x + b > m) or if computing 2x + b
   450  		// overflowed (meaning 2x + b > 2^_W*n > m).
   451  		needSubtraction = not(choice(borrow)) | choice(carry)
   452  	}
   453  	return x.assign(needSubtraction, d)
   454  }
   455  
   456  // Mod calculates out = x mod m.
   457  //
   458  // This works regardless how large the value of x is.
   459  //
   460  // The output will be resized to the size of m and overwritten.
   461  func (out *Nat) Mod(x *Nat, m *Modulus) *Nat {
   462  	out.resetFor(m)
   463  	// Working our way from the most significant to the least significant limb,
   464  	// we can insert each limb at the least significant position, shifting all
   465  	// previous limbs left by _W. This way each limb will get shifted by the
   466  	// correct number of bits. We can insert at least N - 1 limbs without
   467  	// overflowing m. After that, we need to reduce every time we shift.
   468  	i := len(x.limbs) - 1
   469  	// For the first N - 1 limbs we can skip the actual shifting and position
   470  	// them at the shifted position, which starts at min(N - 2, i).
   471  	start := len(m.nat.limbs) - 2
   472  	if i < start {
   473  		start = i
   474  	}
   475  	for j := start; j >= 0; j-- {
   476  		out.limbs[j] = x.limbs[i]
   477  		i--
   478  	}
   479  	// We shift in the remaining limbs, reducing modulo m each time.
   480  	for i >= 0 {
   481  		out.shiftIn(x.limbs[i], m)
   482  		i--
   483  	}
   484  	return out
   485  }
   486  
   487  // ExpandFor ensures x has the right size to work with operations modulo m.
   488  //
   489  // The announced size of x must be smaller than or equal to that of m.
   490  func (x *Nat) ExpandFor(m *Modulus) *Nat {
   491  	return x.expand(len(m.nat.limbs))
   492  }
   493  
   494  // resetFor ensures out has the right size to work with operations modulo m.
   495  //
   496  // out is zeroed and may start at any size.
   497  func (out *Nat) resetFor(m *Modulus) *Nat {
   498  	return out.reset(len(m.nat.limbs))
   499  }
   500  
   501  // maybeSubtractModulus computes x -= m if and only if x >= m or if "always" is yes.
   502  //
   503  // It can be used to reduce modulo m a value up to 2m - 1, which is a common
   504  // range for results computed by higher level operations.
   505  //
   506  // always is usually a carry that indicates that the operation that produced x
   507  // overflowed its size, meaning abstractly x > 2^_W*n > m even if x < m.
   508  //
   509  // x and m operands must have the same announced length.
   510  func (x *Nat) maybeSubtractModulus(always choice, m *Modulus) {
   511  	t := NewNat().set(x)
   512  	underflow := t.sub(m.nat)
   513  	// We keep the result if x - m didn't underflow (meaning x >= m)
   514  	// or if always was set.
   515  	keep := not(choice(underflow)) | choice(always)
   516  	x.assign(keep, t)
   517  }
   518  
   519  // Sub computes x = x - y mod m.
   520  //
   521  // The length of both operands must be the same as the modulus. Both operands
   522  // must already be reduced modulo m.
   523  func (x *Nat) Sub(y *Nat, m *Modulus) *Nat {
   524  	underflow := x.sub(y)
   525  	// If the subtraction underflowed, add m.
   526  	t := NewNat().set(x)
   527  	t.add(m.nat)
   528  	x.assign(choice(underflow), t)
   529  	return x
   530  }
   531  
   532  // Add computes x = x + y mod m.
   533  //
   534  // The length of both operands must be the same as the modulus. Both operands
   535  // must already be reduced modulo m.
   536  func (x *Nat) Add(y *Nat, m *Modulus) *Nat {
   537  	overflow := x.add(y)
   538  	x.maybeSubtractModulus(choice(overflow), m)
   539  	return x
   540  }
   541  
   542  // montgomeryRepresentation calculates x = x * R mod m, with R = 2^(_W * n) and
   543  // n = len(m.nat.limbs).
   544  //
   545  // Faster Montgomery multiplication replaces standard modular multiplication for
   546  // numbers in this representation.
   547  //
   548  // This assumes that x is already reduced mod m.
   549  func (x *Nat) montgomeryRepresentation(m *Modulus) *Nat {
   550  	// A Montgomery multiplication (which computes a * b / R) by R * R works out
   551  	// to a multiplication by R, which takes the value out of the Montgomery domain.
   552  	return x.montgomeryMul(x, m.rr, m)
   553  }
   554  
   555  // montgomeryReduction calculates x = x / R mod m, with R = 2^(_W * n) and
   556  // n = len(m.nat.limbs).
   557  //
   558  // This assumes that x is already reduced mod m.
   559  func (x *Nat) montgomeryReduction(m *Modulus) *Nat {
   560  	// By Montgomery multiplying with 1 not in Montgomery representation, we
   561  	// convert out back from Montgomery representation, because it works out to
   562  	// dividing by R.
   563  	one := NewNat().ExpandFor(m)
   564  	one.limbs[0] = 1
   565  	return x.montgomeryMul(x, one, m)
   566  }
   567  
   568  // montgomeryMul calculates x = a * b / R mod m, with R = 2^(_W * n) and
   569  // n = len(m.nat.limbs), also known as a Montgomery multiplication.
   570  //
   571  // All inputs should be the same length and already reduced modulo m.
   572  // x will be resized to the size of m and overwritten.
   573  func (x *Nat) montgomeryMul(a *Nat, b *Nat, m *Modulus) *Nat {
   574  	n := len(m.nat.limbs)
   575  	mLimbs := m.nat.limbs[:n]
   576  	aLimbs := a.limbs[:n]
   577  	bLimbs := b.limbs[:n]
   578  
   579  	switch n {
   580  	default:
   581  		// Attempt to use a stack-allocated backing array.
   582  		T := make([]uint, 0, preallocLimbs*2)
   583  		if cap(T) < n*2 {
   584  			T = make([]uint, 0, n*2)
   585  		}
   586  		T = T[:n*2]
   587  
   588  		// This loop implements Word-by-Word Montgomery Multiplication, as
   589  		// described in Algorithm 4 (Fig. 3) of "Efficient Software
   590  		// Implementations of Modular Exponentiation" by Shay Gueron
   591  		// [https://eprint.iacr.org/2011/239.pdf].
   592  		var c uint
   593  		for i := 0; i < n; i++ {
   594  			_ = T[n+i] // bounds check elimination hint
   595  
   596  			// Step 1 (T = a × b) is computed as a large pen-and-paper column
   597  			// multiplication of two numbers with n base-2^_W digits. If we just
   598  			// wanted to produce 2n-wide T, we would do
   599  			//
   600  			//   for i := 0; i < n; i++ {
   601  			//       d := bLimbs[i]
   602  			//       T[n+i] = addMulVVW(T[i:n+i], aLimbs, d)
   603  			//   }
   604  			//
   605  			// where d is a digit of the multiplier, T[i:n+i] is the shifted
   606  			// position of the product of that digit, and T[n+i] is the final carry.
   607  			// Note that T[i] isn't modified after processing the i-th digit.
   608  			//
   609  			// Instead of running two loops, one for Step 1 and one for Steps 2–6,
   610  			// the result of Step 1 is computed during the next loop. This is
   611  			// possible because each iteration only uses T[i] in Step 2 and then
   612  			// discards it in Step 6.
   613  			d := bLimbs[i]
   614  			c1 := addMulVVW(T[i:n+i], aLimbs, d)
   615  
   616  			// Step 6 is replaced by shifting the virtual window we operate
   617  			// over: T of the algorithm is T[i:] for us. That means that T1 in
   618  			// Step 2 (T mod 2^_W) is simply T[i]. k0 in Step 3 is our m0inv.
   619  			Y := T[i] * m.m0inv
   620  
   621  			// Step 4 and 5 add Y × m to T, which as mentioned above is stored
   622  			// at T[i:]. The two carries (from a × d and Y × m) are added up in
   623  			// the next word T[n+i], and the carry bit from that addition is
   624  			// brought forward to the next iteration.
   625  			c2 := addMulVVW(T[i:n+i], mLimbs, Y)
   626  			T[n+i], c = bits.Add(c1, c2, c)
   627  		}
   628  
   629  		// Finally for Step 7 we copy the final T window into x, and subtract m
   630  		// if necessary (which as explained in maybeSubtractModulus can be the
   631  		// case both if x >= m, or if x overflowed).
   632  		//
   633  		// The paper suggests in Section 4 that we can do an "Almost Montgomery
   634  		// Multiplication" by subtracting only in the overflow case, but the
   635  		// cost is very similar since the constant time subtraction tells us if
   636  		// x >= m as a side effect, and taking care of the broken invariant is
   637  		// highly undesirable (see https://go.dev/issue/13907).
   638  		copy(x.reset(n).limbs, T[n:])
   639  		x.maybeSubtractModulus(choice(c), m)
   640  
   641  	// The following specialized cases follow the exact same algorithm, but
   642  	// optimized for the sizes most used in RSA. addMulVVW is implemented in
   643  	// assembly with loop unrolling depending on the architecture and bounds
   644  	// checks are removed by the compiler thanks to the constant size.
   645  	case 1024 / _W:
   646  		const n = 1024 / _W // compiler hint
   647  		T := make([]uint, n*2)
   648  		var c uint
   649  		for i := 0; i < n; i++ {
   650  			d := bLimbs[i]
   651  			c1 := addMulVVW1024(&T[i], &aLimbs[0], d)
   652  			Y := T[i] * m.m0inv
   653  			c2 := addMulVVW1024(&T[i], &mLimbs[0], Y)
   654  			T[n+i], c = bits.Add(c1, c2, c)
   655  		}
   656  		copy(x.reset(n).limbs, T[n:])
   657  		x.maybeSubtractModulus(choice(c), m)
   658  
   659  	case 1536 / _W:
   660  		const n = 1536 / _W // compiler hint
   661  		T := make([]uint, n*2)
   662  		var c uint
   663  		for i := 0; i < n; i++ {
   664  			d := bLimbs[i]
   665  			c1 := addMulVVW1536(&T[i], &aLimbs[0], d)
   666  			Y := T[i] * m.m0inv
   667  			c2 := addMulVVW1536(&T[i], &mLimbs[0], Y)
   668  			T[n+i], c = bits.Add(c1, c2, c)
   669  		}
   670  		copy(x.reset(n).limbs, T[n:])
   671  		x.maybeSubtractModulus(choice(c), m)
   672  
   673  	case 2048 / _W:
   674  		const n = 2048 / _W // compiler hint
   675  		T := make([]uint, n*2)
   676  		var c uint
   677  		for i := 0; i < n; i++ {
   678  			d := bLimbs[i]
   679  			c1 := addMulVVW2048(&T[i], &aLimbs[0], d)
   680  			Y := T[i] * m.m0inv
   681  			c2 := addMulVVW2048(&T[i], &mLimbs[0], Y)
   682  			T[n+i], c = bits.Add(c1, c2, c)
   683  		}
   684  		copy(x.reset(n).limbs, T[n:])
   685  		x.maybeSubtractModulus(choice(c), m)
   686  	}
   687  
   688  	return x
   689  }
   690  
   691  // addMulVVW multiplies the multi-word value x by the single-word value y,
   692  // adding the result to the multi-word value z and returning the final carry.
   693  // It can be thought of as one row of a pen-and-paper column multiplication.
   694  func addMulVVW(z, x []uint, y uint) (carry uint) {
   695  	_ = x[len(z)-1] // bounds check elimination hint
   696  	for i := range z {
   697  		hi, lo := bits.Mul(x[i], y)
   698  		lo, c := bits.Add(lo, z[i], 0)
   699  		// We use bits.Add with zero to get an add-with-carry instruction that
   700  		// absorbs the carry from the previous bits.Add.
   701  		hi, _ = bits.Add(hi, 0, c)
   702  		lo, c = bits.Add(lo, carry, 0)
   703  		hi, _ = bits.Add(hi, 0, c)
   704  		carry = hi
   705  		z[i] = lo
   706  	}
   707  	return carry
   708  }
   709  
   710  // Mul calculates x = x * y mod m.
   711  //
   712  // The length of both operands must be the same as the modulus. Both operands
   713  // must already be reduced modulo m.
   714  func (x *Nat) Mul(y *Nat, m *Modulus) *Nat {
   715  	// A Montgomery multiplication by a value out of the Montgomery domain
   716  	// takes the result out of Montgomery representation.
   717  	xR := NewNat().set(x).montgomeryRepresentation(m) // xR = x * R mod m
   718  	return x.montgomeryMul(xR, y, m)                  // x = xR * y / R mod m
   719  }
   720  
   721  // Exp calculates out = x^e mod m.
   722  //
   723  // The exponent e is represented in big-endian order. The output will be resized
   724  // to the size of m and overwritten. x must already be reduced modulo m.
   725  func (out *Nat) Exp(x *Nat, e []byte, m *Modulus) *Nat {
   726  	// We use a 4 bit window. For our RSA workload, 4 bit windows are faster
   727  	// than 2 bit windows, but use an extra 12 nats worth of scratch space.
   728  	// Using bit sizes that don't divide 8 are more complex to implement, but
   729  	// are likely to be more efficient if necessary.
   730  
   731  	table := [(1 << 4) - 1]*Nat{ // table[i] = x ^ (i+1)
   732  		// newNat calls are unrolled so they are allocated on the stack.
   733  		NewNat(), NewNat(), NewNat(), NewNat(), NewNat(),
   734  		NewNat(), NewNat(), NewNat(), NewNat(), NewNat(),
   735  		NewNat(), NewNat(), NewNat(), NewNat(), NewNat(),
   736  	}
   737  	table[0].set(x).montgomeryRepresentation(m)
   738  	for i := 1; i < len(table); i++ {
   739  		table[i].montgomeryMul(table[i-1], table[0], m)
   740  	}
   741  
   742  	out.resetFor(m)
   743  	out.limbs[0] = 1
   744  	out.montgomeryRepresentation(m)
   745  	tmp := NewNat().ExpandFor(m)
   746  	for _, b := range e {
   747  		for _, j := range []int{4, 0} {
   748  			// Square four times. Optimization note: this can be implemented
   749  			// more efficiently than with generic Montgomery multiplication.
   750  			out.montgomeryMul(out, out, m)
   751  			out.montgomeryMul(out, out, m)
   752  			out.montgomeryMul(out, out, m)
   753  			out.montgomeryMul(out, out, m)
   754  
   755  			// Select x^k in constant time from the table.
   756  			k := uint((b >> j) & 0b1111)
   757  			for i := range table {
   758  				tmp.assign(ctEq(k, uint(i+1)), table[i])
   759  			}
   760  
   761  			// Multiply by x^k, discarding the result if k = 0.
   762  			tmp.montgomeryMul(out, tmp, m)
   763  			out.assign(not(ctEq(k, 0)), tmp)
   764  		}
   765  	}
   766  
   767  	return out.montgomeryReduction(m)
   768  }
   769  
   770  // ExpShortVarTime calculates out = x^e mod m.
   771  //
   772  // The output will be resized to the size of m and overwritten. x must already
   773  // be reduced modulo m. This leaks the exponent through timing side-channels.
   774  func (out *Nat) ExpShortVarTime(x *Nat, e uint, m *Modulus) *Nat {
   775  	// For short exponents, precomputing a table and using a window like in Exp
   776  	// doesn't pay off. Instead, we do a simple conditional square-and-multiply
   777  	// chain, skipping the initial run of zeroes.
   778  	xR := NewNat().set(x).montgomeryRepresentation(m)
   779  	out.set(xR)
   780  	for i := bits.UintSize - bitLen(e) + 1; i < bits.UintSize; i++ {
   781  		out.montgomeryMul(out, out, m)
   782  		if k := (e >> (bits.UintSize - i - 1)) & 1; k != 0 {
   783  			out.montgomeryMul(out, xR, m)
   784  		}
   785  	}
   786  	return out.montgomeryReduction(m)
   787  }
   788  

View as plain text